Long-term outdoor operation of wind turbines tends to cause multiple concurrent faults, while maintenance crews, working hours, and spare-parts inventories remain limited. Manual maintenance scheduling and traditional optimization approaches have difficulty coordinating strongly coupled tasks and resources, resulting in severe power losses and resource conflicts. This paper proposes a deep reinforcement learning-guided Hybrid Evolutionary Framework (DRL-HEF) to improve scheduling efficiency and resource coordination. First, a tri-objective scheduling model that considers power-generation loss, risk waiting time, and switching cost is designed, together with crew availability, working-time windows, and spare-parts circulation. Second, maintenance resources and historical fault information are utilized to train a Hierarchical Transformer-based Multi-value Decomposition Proximal Policy Optimization (HT-MDPPO) policy, which learns task-crew assignment knowledge and generates feasible schedules for Non -dominated Sorting Genetic Algorithm II (NSGA-II) population initialization. Finally, an hypervolume-stagnation-triggered feedback dynamically injects schedules which are generated by HT-MDPPO into the population to overcome search stagnation. Experimental results demonstrate that DRL-HEF accelerates evolutionary convergence while effectively coordinating crew deployment and spare-parts allocation, thereby reducing maintenance conflicts and generating reliable maintenance schedules for large-scale onshore wind farms.
 

onshore wind farms,maintenance scheduling,deep reinforcement learning,hybrid evolutionary framework